Machine Learning Leveraged Simulation Digital-twins Enabling Real-time Detailed Insights


Jenil Dedhia (jdedhia@anagha.consulting), Anagha Consultants

https://www.prodtwin.com/

https://www.anagha.consulting/


Leveraging data-based/ machine learning approaches for predictive capabilities is a common theme currently in the various stages of the product/ process development life cycle. This helps with linking the relevant inputs of a process to the outputs, without necessarily understanding the underlying physics or interactions.

Figure 1

When one has enough experimental data (either from prior runs or structured analysis) on various inputs and outputs for a given process comprising a set of unit operations, such approaches can be used to develop a process model (Figure 1). Within the applicable parameter window, treating the whole process as a black box, such a model provides predictability into the outputs because of changes in any inputs. We will talk more about this in a separate post.

Focus of this post

In the specialty materials industry (like in pharma or niche materials), where getting experimental data both at the lab or a large scale can be prohibitive, physics-based process models can provide the necessary initial guidance to the practitioner for decision making. Depending on the complexity of the physics involved, such process models may require advanced computations and simulations involving computational fluid dynamics (CFD) or Discrete Element Modeling (DEM). Even with increased computational power, such process models may not meet the needs of a practitioner to get visibility and predictability for real-time decision making. With example cases, this post focuses on leveraging machine learning approaches to address such practitioner needs in providing real time guidance grounded on process physics. (Figure 2)

Figure 2

What is a Simulation?

In process industries, simulation software is traditionally used to obtain insights on the process or asset without having to do experiments. Simulations are typically used in process design and optimization to predict what may happen in the real world and run ‘what-if’ scenarios.

Why ML-leveraged simulation twins?

Detailed simulations which capture the process reasonably are generally computationally-intensive and hence time-consuming, sometimes taking hours to days to give results. This has been hampering the adoption of these tools, like CFD and DEM. AI/ML techniques can be applied on simulation data to provide the practitioner on-the-fly detailed-level insights obtained from the simulation. This enables the practitioner to play out extensive ‘what-if’ scenarios to gain detailed process understanding and to leverage such insights for real-time decision-making.

Let us look at a couple of scenarios.

Use Case (I): ‘Forward’ model

Here we present a case where the developed ML-twin can predict the outcome of a simulation by taking in the inputs which go into the simulation model.

Below is an example of DEM simulation of Powder Flowability test [1], where a bulk powder property Static Angle of Repose (SAoR) is obtained as an outcome of the simulation as shown in Figure 3. Simulations can be run over a parametric space of inputs (like coefficient of static and rolling friction, Young’s Modulus, particle size, particle cohesiveness) to generate data for the ML-model. Here, we use data of 53 simulations reported in the literature [1].

Figure 3

Once the simulation data is cleaned and pre-processed, different machine learning techniques (like Linear Regression, Random Forests, Gradient Boosting, and even Neural Networks) can be evaluated and the best model is identified. Figure 2 shows the results obtained for the model developed to predict SAoR through the above data. The accuracy of the test results indicate no significant dilution of the quality of the SAoR predicted via machine learning approach compared to the detailed simulation.

Figure 4

Hence, an initial investment of some simulations (in this case total ~300 hours) can be used to build powerful ML models which can eliminate lot of simulations (in this case, each ranging from few minutes to few hours) in the future and give real-time insights to the users.

Use-Case (II): ‘Reverse’ model

Here, we present a case where the developed ML-twin can predict one of the inputs needed for the simulation for a given desired simulation outcome.

The classic case for such a model is calibration of DEM material parameters. DEM needs intrinsic powder properties as inputs, some of which are hard-to-measure (like for example particle-particle friction properties) for practitioners in the pharma or the additive manufacturing space. This limits the value add of such modeling approach for an industrial practitioner. One way this issue is currently addressed in this field, is as follows [2]. As many bulk properties of the powder (like bulk density, SAoR, dynamic angle of repose etc.) are easier to measure, the practitioner is advised to run multiple DEM simulations of these bulk property setups to identify the right combination of intrinsic properties.


Readers who are acquainted with DEM simulations would know this is a tiring and time-consuming exercise, requiring lot of iterations of simulations. Figure 5 depicts how ML-models can be leveraged to eliminate the calibration exercise by predicting the input property of interest for a desired simulation outcome, in this case, experimentally obtained SAoR.

Figure 5

Key Take-Aways

      • Detailed simulation techniques like DEM, CFD provide critical insights on the process. They require expert users and take long turnaround time for providing insights. This makes them less attractive to the practitioners interested in real time decision making.

      • Advanced artificial intelligence techniques are now available which can be leveraged using these simulation data to provide on-the-fly guidance for the practitioners to take real-time decisions in their routine efforts. As an example, how manual and cumbersome steps within two use cases (involving the application of DEM approaches, and calibration of material parameters) can be expedited, are presented.

      • These ML-models can be ‘Forward’ or ‘Reverse’, i.e. can predict the simulation outcome or the input parameters for obtaining a desired given simulation outcome.

      • Such approaches can help practitioners expedite their process development journey through process understanding.


We hope this article helped you understand how data-driven approaches can be combined with advanced simulation techniques to provide practitioner friendly and simulation level detailed visibility into the process. Such approaches can be leveraged over a range of complex applications, across a range of industries.

See value in this for your process needs?

Interested in exploring how to tailor such a hybrid approach to expedite your process development?

Reach us out at Jdedhia@anagha.consulting


PS: These use-cases were developed on ProDTwin platform (www.prodtwin.com), an offering of Anagha.


References

[1]. El-Kassem, B., et al, 2020, Computational Particle Mechanics https://doi.org/10.1007/s40571-020-00315-8

[2]. https://rocky.esss.co/new/rocky-dem-launches-calibration-suite-for-accurate-parameter-setting/